Why a random control group matters – and why one alone is not enough

A control group is intended to answer a simple question: How does the same test logic behave on artificially generated price paths whose structure is controlled rather than taken from observed market history? This comparison is especially important when a trend-following decision logic produces similar results across many real instruments and variants. Without a control, it remains possible that the testing architecture itself creates a particular pattern.

What Version 1 did

The real research universe contained 94 market instruments selected for the test matrix. Six separately reported synthetic control series were added, creating a uniform matrix of 100 price series: 94 real instruments + 6 synthetic control series. Across 60 oscillators this produces 5,640 real-instrument tests and 360 control tests, or 6,000 test combinations in total. The synthetic series are explicitly not counted as real markets.

The controls are not six identical random-walk copies. They were generated as independent synthetic OHLC price paths using different parameter sets. The generator varies, in particular, starting drift, trend-switch probability, drift persistence, drift noise, close noise and additional high/low range. The result is a set of artificial paths with materially different degrees of trend persistence. This is deliberate because the Version 1 logic is trend-following rather than a mean-reversion model.

This is useful evidence that real and artificial data behave differently under the current test logic. It is not, however, a universal rejection of every null hypothesis.

The six synthetic trend regimes

The generator workbook describes the stored series using an Efficiency Ratio: absolute net change in the close series divided by the sum of absolute close-to-close changes. Values near 0 indicate a strongly meandering path; values closer to 1 indicate a much more directionally efficient trend path.

Why six trend regimes are still not enough

The six series differ substantially, but they all come from the same generator family. They therefore test several degrees of synthetic trend persistence, not every possible null model. Artificial comparison data can also be constructed to preserve parts of observed return distributions, short-term dependence, volatility clustering or other market properties. The question answered by a control depends on which structures are preserved and which are removed.

What Version 2 still needs to make reproducible

The existing workbook documents the generator logic, parameter fields and the six frozen control series. The generation process itself uses Excel random functions. For independent regeneration of identical paths, Version 2 should therefore use a deterministically reproducible generator with a documented seed or an equivalent reproducible random source. Distribution, autocorrelation, scaling and the treatment of volatility and gaps should also be documented formally. The current control should therefore be treated as a documented comparison result, not as the endpoint of the statistical analysis.

Why this makes the claim more precise rather than weaker

The correct conclusion is not “the control group proves the market edge”. A more precise statement is: Under the Version 1 control design, the real test matrix separates clearly from six parametrically different synthetic trend regimes. Version 2 is intended to test whether that separation persists when additional, differently constructed null models are generated reproducibly.